Quick answer

Read GPU hotspot as a delta, not a single number. Compare it with the edge temperature from the same sustained run, confirm your tool version supports the sensor, and check the card’s thermal-slowdown flag. A wide delta that persists after a correct remount points at paste or mounting, not the cooler design.

Contents

At a glance

Class
PC Hardware

Provenance

  • Nov 16, 2026Published

A reader sends me an overlay screenshot about once a week. The GPU temperature line says something calm, the hotspot line says something alarming, and the question is always the same: is the card dying? Almost every time, the honest answer is that one number cannot tell you, because the useful measurement is the gap between the two.

I am not going to hand you a magic delta threshold. Neither NVIDIA nor AMD publishes one, and any site that gives you a single number is guessing on your behalf. What the vendors do document is how to read the sensors and how to tell whether the card is actually being slowed down by heat. That is enough to make a real decision.

What the hotspot reading actually is

The hotspot, which AMD’s software calls the junction temperature, is the hottest point on the graphics die. The ordinary GPU temperature line is a different, lower sensor, usually described as edge or core temperature. Our own GPU hotspot delta census defines it plainly: the hotspot is the highest reading among the temperature sensors on the die, and the delta is hotspot minus core from the same session.

That definition matters because it tells you what the delta is measuring. It is not a verdict. It is a description of how evenly heat is crossing from silicon into the cold plate. A wide spread means one region of the die is much hotter than the part the edge sensor sees.

NVIDIA does not expose the field in its own command-line tool. The nvidia-smi documentation describes board temperature readings and a T.Limit sensor that reports the margin to the maximum operating temperature, and it does not list a hotspot reading anywhere. On RTX 50 cards the number comes from third-party software, which is exactly why tool version matters so much here.

Edge, hotspot, and memory are three different jobs

People collapse these into one “GPU temp” and then argue about it. Keep them apart:

ReadingWhat it describesWhat it is good for
GPU temperature (edge)A core sensor, typically the cooler-running of the die readingsTracking overall cooler performance and fan response
Hotspot / junctionThe hottest die sensorJudging how evenly the die is being cooled
Memory junctionGDDR package sensorsJudging memory cooling, separate from die contact

The delta is computed from the first two, and only if both come from the same sustained run. Our census validator rejects a hotspot below core, a hotspot outside its documented 20–125 °C range, a core outside 20–100 °C, and any load shorter than ten minutes. Those rules exist because a ten-second screenshot of a menu is not a thermal measurement.

This is the single most common mistake I see in forum threads: someone records a core maximum from one run and a hotspot maximum from another, subtracts them, and posts the result. Two session maxima are not a delta. They may have happened minutes apart under different loads.

The bands, and what I check in each

Our census publishes exact-model distributions in five fixed bands rather than inventing a universal limit: 0–9 °C, 10–19 °C, 20–29 °C, 30–39 °C, and 40 °C or higher. The edges are part of the dataset version, so a row collected this year stays comparable to one collected next year.

Here is how I use each band. This is a diagnostic posture, not a measurement I am claiming to have taken:

Delta bandHow I read itFirst thing I check
0–9 °CUnusually tight. Possible, but confirm the tool is not repeating one sensor.Tool name and version; whether the hotspot field is supported on this card
10–19 °CThe ordinary working range for a card with good contactNothing, if clocks and throttle flags are healthy
20–29 °CWorth watching. Often load, airflow, or fan-curve dependentSustained-load duration, case airflow, fan curve
30–39 °CA contact or mounting conversation is reasonableMounting pressure, paste or pad state, cooler seating
40 °C or higherThe pattern I treat as a fault until proven otherwiseFull remount, or warranty service if the card is covered

Two cautions before you act on that table. A zero-degree delta is valid in our schema, but a tool that is reading the wrong sensor can also produce one, so verify support before you celebrate. And a delta in the 20s on a hot day in a cramped case is not the same finding as the same delta in a ventilated tower at stock settings.

Confirm the tool before you believe the number

This is the step almost everyone skips, and it is where half the “my hotspot is broken” reports come from.

On the NVIDIA side, CPUID’s HWMonitor changelog records hotspot temperature on RTX 50 GPUs as fixed in version 1.65.1, and version 1.66 as improving the hotspot report on AMD Radeon RX 9000 cards. On Linux, LACT v0.10.0 says the hotspot sensor on Blackwell “now works with a Blackwell-specific path, instead of reporting bogus values,” and the project’s pull request 1122 is where that path was added.

“Instead of reporting bogus values” is the sentence to remember. A wrong sensor path does not usually throw an error. It hands you a plausible number that describes something else, and you build a whole theory on top of it.

On AMD cards the driver exposes both readings directly. AMD’s performance metrics article DH3-038 lists both GPU Current Temperature and GPU Junction Temperature among the metrics you can log, with a sampling interval adjustable from 0.25 to 5 seconds. DH3-020 notes that GPU Hotspot Temperature is only available on Radeon VII, Radeon RX 5000 Series and newer products. If you are on an older card, the field may simply not exist on your hardware.

So the first check is free and takes ten seconds: write down the tool name and exact version next to every reading you plan to compare. That record survives a reboot, and it is the one piece of evidence you will still have when the numbers stop making sense.

Is the card actually being slowed down?

A wide delta is not proof of throttling. It is proof of an uneven thermal profile. Whether that costs you performance is a separate question, and both vendors expose a flag for it.

On NVIDIA hardware, the NVML clocks event reasons separate software and hardware thermal slowdown from power capping and other reasons for reduced clocks. In practice that means you can see whether the card is being held back by heat or by its wattage limit, which are completely different problems with completely different fixes.

The T.Limit reading in nvidia-smi is the quick version. It reports margin, not temperature, and the documentation says a T.Limit value of 0 °C or lower indicates the GPU may optimize its clock based on thermal conditions. A healthy margin that stays positive under load is a good sign.

Run the query while the load is running, not after:

nvidia-smi -q -d TEMPERATURE,POWER,CLOCK,PERFORMANCE

Save that output to a file. A log on disk survives a reboot; a live overlay reading does not, and the moment you restart to reseat a cable you lose the evidence you were about to reason from.

On AMD, the junction reading and the clock behavior in the same overlay give you the same comparison, with the 0.25-second sampling floor from DH3-038 fine enough to catch a spike.

Ordered checks, cheapest first

Work down this list and stop as soon as the pattern resolves. Each step is cheaper and less invasive than the one below it.

  1. Take one honest reading. Same session, same load, at least ten minutes, both numbers recorded together. If you cannot satisfy that, you do not have data yet.
  2. Verify the tool version. Check the changelog for your card family. On RTX 50, HWMonitor below 1.65.1 or LACT below 0.10.0 is not a reading you should build a decision on.
  3. Check the throttle flag. Positive T.Limit margin and no thermal slowdown event means the delta is not costing you clocks, whatever it looks like.
  4. Look at fan behavior. A curve that never ramps, or a zero-RPM mode that never exits under load, produces a wide delta for reasons that have nothing to do with paste. The GPU fans stuck at full speed guide covers the opposite failure, where the controller reacts to a sensor you cannot see.
  5. Test case airflow without changing hardware. Remove the side panel while the load runs and watch whether both readings fall together. If they do, the problem is the enclosure and the fan curves, not the card.
  6. Only then consider the mount. If the delta stays wide with airflow and fans sorted, the remaining suspects are mounting pressure and the thermal interface. Even screw torque in an X pattern, verified contact, and a fresh interface are the correct sequence. Phase-change pads exist precisely because ordinary paste can migrate out of a bare-die joint over many heat cycles.
  7. Undervolt last. Once the mechanical side is right, a flatter voltage-frequency curve reduces total heat. It does not repair contact, and applying it first just moves both numbers down and hides the fault.

If you want the full flag-by-flag walkthrough for both processors and graphics cards, the thermal throttling fix guide goes deeper on the telemetry side, and the GPU overheating guide covers the physical checks. For a model-by-model ceiling reference, the safe GPU temperature guide is the companion piece.

When to stop

Stop diagnosing and start deciding when one of these is true.

  • The card is under warranty and the delta pattern points at contact. A support ticket is cheaper than a voided warranty. Policies vary by brand and region, so read the terms for your exact model before the screwdriver comes out.
  • The delta stays wide after a correct remount. One careful remount with verified contact and an even torque pattern is a reasonable attempt. A second and third are not diagnosis, they are hope. If the spread persists, the card needs service or replacement.
  • The hotspot reads below the edge temperature. That is not a hot card, that is a sensor or tool problem. Recheck the version before you touch hardware.
  • The throttle flag never fires and clocks hold. Then you have a cosmetic number, not a fault. Close the overlay and go use the machine.

If none of the checks resolve it, the diagnostic workbench lets you record each test result and the readings beside it so a technician gets a record instead of a story. And if you have a reading worth contributing, the census intake is open — that dataset is how the bands above get better over time.

Frequently asked questions

What is a GPU hotspot temperature?
The hotspot, also called the junction temperature, is the highest temperature reported among the thermal sensors spread across the GPU die. The ordinary GPU temperature reading is a separate, lower sensor usually described as edge or core temperature. The distance between the two is the hotspot delta, and that delta is more informative than either number alone because it describes how well heat moves from the silicon into the cooler.
Is a 100 C GPU hotspot bad?
Not by itself. A high hotspot is only meaningful next to the edge temperature, the workload, and the card’s thermal-slowdown flag. If the edge reading is close and clocks hold above the rated base frequency, the card is boosting as designed. If the edge reading is far lower, fans are at maximum, and the throttle flag is active, the delta is telling you heat is not crossing into the cold plate evenly.
Why does nvidia-smi not show a hotspot temperature?
NVIDIA’s nvidia-smi documentation describes board temperature sensors and a T.Limit margin to the maximum operating temperature, but it does not list a hotspot field. On RTX 50 cards, hotspot readings come from third-party tools. CPUID’s HWMonitor changelog says hotspot temperature on NVIDIA RTX 50 GPUs was fixed in version 1.65.1, and LACT’s v0.10.0 release notes describe a Blackwell-specific sensor path.
What core-to-hotspot delta means the paste has pumped out?
There is no universal cutoff, and no vendor publishes one. What matters is the combination: a wide spread that stays wide across repeated sustained runs, fans pinned near maximum, and the thermal-slowdown flag active while clocks fall. If that pattern survives a correct remount with even mounting pressure, the paste layer or pad is the remaining suspect rather than the cooler’s contact with the die.
Does undervolting lower the hotspot temperature?
Undervolting lowers power, so both edge and hotspot temperatures fall, but it does not fix a contact problem. If the spread between the two readings is wide, lowering the total heat only moves both numbers down together and hides the fault. Fix airflow, fan behavior, and mounting first, then tune the voltage-frequency curve once the delta looks sane.
Can a hotspot reading be wrong?
Yes, and that is why the tool and version matter. LACT’s release notes say the Blackwell hotspot sensor previously reported bogus values until a Blackwell-specific path was added in v0.10.0. A reading from an unsupported or outdated tool can look plausible while describing a different sensor entirely, so record the tool name and version beside every reading you compare.
Should I repaste a graphics card under warranty?
Check the card maker’s warranty terms before you open anything. Disassembly can affect coverage depending on the brand and region, and a card that is still under warranty may be better served by a support ticket than by a screwdriver. If the card is out of warranty and the delta pattern points at contact, a careful remount with the correct thermal interface is the sensible next step.
Do memory temperatures matter as much as the hotspot?
They measure different things. The hotspot tracks the graphics die; memory junction sensors track the GDDR packages. A card can have a healthy die delta and still run hot memory, which is common on cards with dense GDDR6X or GDDR7 arrays. Watch both, but do not treat a memory reading as evidence about die mounting.

Sources and corrections

Last updated
Methodology
See our methodology for research and review standards. It draws on 8 cited sources, listed below, each checked against the original page on the date above. Sensor definitions, telemetry fields, and tool version behaviour verified November 16, 2026 against NVIDIA’s nvidia-smi and NVML documentation, AMD’s Adrenalin performance-metrics and performance-tuning support articles, CPUID’s HWMonitor changelog, and the LACT release notes and pull request that added Blackwell hotspot support. Distribution bands and sampling rules are this site’s own census definition, linked in the body. I did not measure a card for this page and quote no temperatures as mine.
Corrections
Spotted an error? Email a correction with the affected line and what you found. The draft includes this page’s address. Verified changes are recorded here. Contact details.

About the author

Written by Lowell K. Wood IV, who builds and runs TechFuelHQ from St. Louis, Missouri.